this is really helpful for me. thx a lot.

AttributeError: 'BehaviorSpec' object has no attribute 'observation_shapes',how to solve it?

can i also implement Deep Q Network for training the agents?

Screen with game shows up after env=.... But command prompt doesn't let me write code as well as the popped up game doesn't five me wait button like it did to u

can i using Deep Q Network for training the agents? example for create a agent AI Car for racing game

Awesome!! Really Great content. I’ve learned a lot, thanks for sharing

Thank you so much for this. Fascinated by machine learning and unity-ml seems like a great place to explore. You've saved me a lot of headaches already and I can't wait to dig through the rest of your videos.

Help me, console said Cant Connect on port 5004, and the command shell said " AttributeError: 'str' object has no attribute '_key' " please help me been stuck on this for 4 days straight cant find best solution

"After a bit of crying..." So on point... This tutorial is superb! Thank you so much for the time and effort you have put into making this video, it is immensely helpful!

you didn't explain selling at market value, selling at market doesn't sell at last traded price it sells within X% of where it's at, could be higher could be lower than what the price was when you executed market price bell/buy. for such a long video I was really hoping that you were going to get into everything, like the order book, ADL, index prices. what everything on the screen means and how it all comes into play with each other.

Great run-through. Appreciate you touching on Auto-Deposit Margin, was wondering about that.

This is the awesome lecture I've ever seen. Thanks to you I could handle the ML-Agent 2.0. Especially I got hard time with "Action Buffers and Heuristic parts". I'm looking forward to the next lecture on "Discrete Action". Thank you Thank you again for your painful effort and a great lecture!

Awesome explanation, wonderful video, thank you

Hi, this is a great tutorial but Im really confused about one thing. I'm running into an issue where my agent won't train properly. The game is very similar to this one but slightly different I have an arena with 4 walls where I have a character. Both the agent and an apple spawn randomly. The entire goal of the game is for the agent to run into the apple and collect it (just like this game). I have practically the same observations & code etc as you, but it doesn't seem to learn, and instead just insists on running in a random direction instead of learning to use the coordinates its given for the apple. I thought "Observable Attribute Handling" might be the culprit but I see you have it on Ignore as well. The only thing different about my game is that it doesn't reset after the agent picks up the apple - everything's handled from a controller script which resets the game every 10 seconds, no matter the state. I've tried it with/without that and it doesn't seem to make a difference, however. The full list of observations I'm passing are (7 total): - Apple count within radius (of entire map for testing purpose) - Nearest apple vector3 (there's only one) - Agent's own vector3 Inputs are exact same as yours (2 continuous), except using a controller.Move instead of rigidbody. Code: - Agent Movement: www.toptal.com/developers/hastebin/ocodezijon.csharp - Spawner: www.toptal.com/developers/hastebin/usuzakehoc.csharp Any ideas? Here's a GIF of it: gyazo.com/29c37778c005b6fafd49262cbd1b45ac , this is after 1000+ gens. I ran it overnight in a similar setting with 5million steps and it still wasn't working. I suppose the only real difference is the chance of random collision into the apple to begin with is much smaller due to the gameplay space? Trying it with a large apple but the optimal strategy seems to be the same of just running in a random direction, and not using any of the observation date: gyazo.com/5c1cf2abb3d0d52dc4dc8a06fd6bea48 thanks sorry for the long comment

Hi! Thanks so much for this tutorial, so so helpful! I've been applying it to my own custom environment. I've got everything working the way I want it except that for some reason when I reset after the first episode, I cannot move my player anymore, and the same happens when the agent is training... Any idea why that might be happening? Thanks in advance 🙏🏽

This was so informative ! Thanks a ton for making this 🤩. Just wish I came across this a few months earlier so that I wouldn't have to struggle to figure stuff out on my own - would have saved a ton of time & energy. Thanks once again ! 🎉

Hi, quick question. How can I deal with the "Academy is inaccessible due to its protection level c#" error when I inherit the class?

A truly enjoyable tutorial. Good length for the depth of information. Really well presented with all the needed details articulated clearly and concisely in a pleasant and friendly manner. You have a gift for teaching. Thanks!

Very informative video! Thank you.

Awesome video, thank you !

Bit Corn lol!

I tried this and had to change a couple of lines in the agent.cs: I believe is OnActionReceived(ActionBuffers actionBuffers) and Heuristic(in ActionBuffers actionsOut) instead of float[] lists

Very well presented - natural and honest. Thank you for taking the time to help the community. It is appreciated.

That was really cool, and rather inspiring on the application of machine learning for Protein Analysis. Makes me think. I do however have a couple of queries: a) It sorta feels as though train_data still has the label indicator values left in the final column 1000, so the Nueral net may be seeing that last feature, and possibly overfitting against that when I remove that last column in val_data, I get the validation error of around 6%, not that different to Logistic Regression, and then ROC_AUC of 0.69, because of an excess of false Negatives (0s) b) I thought in pytorch we were supposed to call optimizer.zero_grad(), before doing the loss.backward calls ? However when I do put this call in, I could not get my network to consistently train, improve Anyhow I found this really interesting stuff.

One of the friendliest and most honest walk-through tutes on this rapidly changing and, at times, frustrating ml-agents environment. Enjoyed it so much!!! well done and thank you.

That kinda useful and informative. The ML-Agents examples tend to use Multiple parallel Environments for faster Training. Will this Build and execute process against a Unity project Environment which has multiple copies of the Environment/Agent instances For faster Training ? I presume that it would recognise that ?

This is really really cool. I have been meaning to try adding some learned AI to a few unity projects and watching this is a huge motivator! It seems like there is a lot to know, and not really having any idea how much learning I have yet to do makes it intimidating - at least for me. I didn't want to start with something that was outdated, so I was happy when I saw this v2.0 video. I appreciate it!

Thank you so much for this. I appreciate you and this so much. This helps set up a whole foundation that I'll need. I also love how intuitive and relatable you make the whole process and how clear you make things on parameters that need to be set. Computer science is a male dominated space and it's refreshing to see more women becoming more included and involved in the CS community. You've made this whole ML-Agent experience that much more accessible and I'm very grateful for that.

I get you but your screen is to small can't really see Mam

That was going great Rachel - until we got the bit about the Trading Password. If this is a beginners guide you needed to have included that in the instructions (bc that's where I clicked away) Personally I was looking for up to date info on funding with Fiat - I'm not going to set up on another Ex just to buy BTC when I can fund direct.. I hope this helps you structure future content.

thank you so much! helps a lot!

Thanks babe

This video is BLANK but does have audio.

this is the most useful guide on kucoin I have seen on youtube even though I lost my entire investment

I have problem with "float[]" thing, i copied code from github and it gives this message "Assets\agent.cs(69,26): error CS0115: 'RollerAgent.Heuristic(float[])': no suitable method found to override"

Hi Rachel. Thank you so much for the tutorial. When I clicked the Play button for the first time, the ball does not seem to respond to my keyboard input. The console shows me the following message: "Couldn't connect to trainer on port 5004 using API version 1.0.0. Will perform inference instead." Is there a way for me to fix this?

Great stuff, super helpful to crack the code!

Versions compatibility issues is exactly what has kept me away from playing more with ml agents to date.

Oh my, oh my, so precise and to the point. Great work and keep it coming! 🤩

gracias a tu video pude llegar mas de la mitad :) pero me tranque en el minuto 21:00 :( pero gracias :,)

Is it possible to use ML-Agents to create strategic AI that targets specific individuals, such as detecting the habits of players and exploiting their weaknesses in 3D action games?

Thank , can i use it for enemies ai ?

Appreciate for the tutorial video! I got a question about code in expected_sarsa.py line 41. I have changed this code below: action = env.behavior_specs[behavior_name].create_random_action(len(decision_steps)) to action = env.behavior_specs[behavior_name].action_spec.random_action(len(decision_steps)) since there is some of updates in base_env.py ... But this code occurs this error below and I can't find solutions.. : IndexError: only integers, slices (`:`), ellipsis (`...`), numpy.newaxis (`None`) and integer or boolean arrays are valid indices Is there any solution?

It doesn't seem to work because of the mlagents version I guess. I have the same Unity version (2019.4) and mlagents package (1.5.0) but I'm wondering if the reason some functions do not exist (for example is_action_continuous() ) is because of the python mlagents version? I installed in python mlagents and mlagents-envs version 0.27.0

This was great. Glad I stumbled upon your channel. Looking forward to follow into the future 👍😊😎

drive.google.com/file/d/1NjMbSSbooId82Mu1XU0lcyUw19VGHfvN/view?usp=sharing

Thank you for ml-agents video. For me Reinforcement with ml-agents & Unity seems to be hard to get clear idea how to make own working projects etc. There is couple of books from ml-agents, but all books seems to be quite old. Best source so far is KZbin.

FYI volume is WAY too quiet. Thanks for the video.

I am so grateful that this video exists! Thank you!

I followed every step exactly and when I tested this by using 'Heuristic only' behavior type, it worked. But When I changed it to 'Default' and enter the command "mlagents-learn config/Roller_Ball_tutorial_config.yaml --run-id=RollerBallTest --force" and press the play button, the RollerAgent does not move.. Is there any solution?